AI answersn 17,262window Aug 7 to Oct 20, 2026questions 1315 enginescited pages read 2,398

Links get you cited. Mentions get you named.

Between Aug 7 and Oct 20, 2026, sourcesignal’s free report asked 131 buying questions on five engines (ChatGPT, Gemini, Claude, Google AI Mode and AI Overviews), 17,262 answers in all. We fetched the 3,179 pages that carry four fifths of their citations, read the 2,398 a plain request could open, counted the brands each page names and links to, and looked up the domain authority of 580 brand websites. The engines name the brands their sources name. A brand no cited page names is named in 8.5% of answers; a brand seven or more cited pages name, in 74.6%. Two in three mentions on the cited pages carry no link to the brand. Inside one question on one engine, the mentions predicted naming better than domain authority 234 times in 282. Authority helps a little, and most of that travels with being written about.

01 / the old currency

In search, links were one signal of a few, and domain authority the weakest of the backlink measures.

Search ranks pages, and links were one of the ways it judged them. Across the top 20 results for 1,000,000 US keywords, Ahrefs found the number of sites linking to a page correlates 0.255 with its position and Domain Rating 0.131, which Ahrefs itself calls “a small positive correlation, and far from definitive” and says Google does not use. Google’s engineers described its core ranking in court as anchors (links), body (content) and clicks. Its spam policy called links “an important factor” until March 2024 and “a factor” since, and its own description of quality asks whether prominent sites “link or refer to the content”. Links counted in search, and they were never the whole ranking.

Page referring domains: 0.255
Followed referring domains: 0.250
Backlinks: 0.248
Domain Rating: 0.131

Correlation (Spearman) between each measure and a page’s position in Google, the top 20 results for 1,000,000 US keywords: Ahrefs, Jan 30, 2025. 1 would be the same order; 0, no relation.

n = 1,000,000 US keywords, top 20 results eachSpearman with position inside each results pageDomain Rating is Ahrefs’ link-based score, 0 to 100no number in this section comes from sourcesignal’s data

02 / the engine names who its sources name

The more of a question’s cited pages name a brand, the more often the engine names it: from under one answer in ten to three in four.

No cited page names it: 8.5%, 5,338 cases
1 cited page: 25.9%, 1,141 cases
2 to 3 cited pages: 38.3%, 787 cases
4 to 6 cited pages: 52.9%, 330 cases
7 or more cited pages: 74.6%, 139 cases

Share of answers that name the brand, by how many of the batch’s cited pages (the ones we could read) name it, pooled over 7,735 cases in 449 batches. A brand’s own site never counts as a page naming it.

n = 7,735 cases, 449 batches, 57 reportsnamed: the brand’s name as a whole word, any casecited pages: the 2,398 we could readassociation: retrieval and naming can share a cause

An engine that searches reads a handful of pages and writes its answer from them. A brand no cited page names is named in 8.5% of answers, a brand one page names in 25.9%, and a brand seven or more pages name in 74.6%. Inside one batch, the brands the pages name most are the brands the engine names most: a median correlation of 0.58, positive in 307 of 329 batches and for all 41 businesses with enough data. Holding the brand fixed, so its authority and fame stay the same, naming still moves with what each question’s pages say (0.49, positive for 464 of 573 brands). The businesses and the competitors they typed in, brands nobody selected for being named, climb the same curve from 14.5% to 89.4%. Who gets named tracks the pages the engine cites.

03 / backlinks are not mentions

Two of every three times a page AI engines cite names a brand, it does not link to it.

A link and a mention are different marks on a page. On the cited pages we read there are 2,808 mentions of a brand with a known website (a page naming a brand counts once), and 1,877 of them (66.8%) carry no link to the brand’s site. In 43.8% of batches no cited page we read links any of the brands tested; in 23.1%, none names one. The reverse almost never happens: a brand linked from a cited page and named on none turned up in 40 of 5,364 cases. Best-of lists take a quarter of all citations (25.2%), from one cited page in six. The answers do the same: 49.6% of the answers that name the business give no link to its site. Counting links misses two mentions in three.

Named on a cited page, no link to the brand
Every mention of a brand with a known site: 66.8%, 1,877 of 2,808
The same, weighted by how often each page is cited: 65.4%, 38,845 of 59,430
Brands nobody selected for being named: 66.7%, 342 of 513
Linked on a cited page, never named
Linked from a cited page, named on none: 0.7%, 40 of 5,364 cases

On the 2,398 cited pages we could read: the share of a brand’s mentions with no link to its site or a subdomain, and the share of cases where a cited page links the brand but none names it. Brands with a known site only.

The 1,877 mentions with no link to the brand, by what the name does on the page.

n = 2,808 page-brand mentions, 5,364 casesa link built by a script is invisible to a plain requestan affiliate redirect counts as no linka brand’s own site never countsbest-of list: best or top in the title or address

04 / mentions beat links

On the same brands, the pages’ mentions predict who is named better than their links do: in 131 batches to 53.

Every brand with a known site22 ties: links predict better 53; mentions predict better 131
Counting batches with no brand linked24 ties: links predict better 64; mentions predict better 195
Brands nobody selected17 ties: links predict better 5; mentions predict better 17

Batches where each measure follows naming more closely (a higher rank correlation), the same brands on both sides; ties left out. The middle row counts batches where no brand is linked as a win for mentions when mentions follow naming.

Mentions on cited pages: 0.44, 0.40 to 0.49
Links from cited pages: 0.14, 0.10 to 0.19
Domain authority: 0.05, 0.02 to 0.08

Each measure’s weight in one model of naming with all three together, on ranks inside each batch: 358 batches, 5,167 cases. Beside each, its 95% interval.

n = 206 batches where both measures are defined; 283 counting the restlinks: pages linking to the brand’s site or a subdomainmodel: ranks within each batch, intervals from 2,000 resamples of whole batches

A link almost always comes with the name, and the name usually comes without the link. Compared on the same brands, the pages that name a brand follow who the engine names more closely than the pages that link to it in 131 batches, and less closely in 53 (22 ties), a median correlation of 0.61 against 0.46. Counting the batches where no brand is linked at all, mentions win 195 to 64. In one model holding mentions, links and domain authority together, mentions keep a weight of 0.44, links 0.14 and authority 0.05. In a model of mentions and links alone for the businesses and the competitors they typed in, mentions keep 0.56 and the links’ 0.11 cannot be told apart from zero. A link adds a little; the mention carries most of the signal.

05 / authority mostly travels with mentions

Authority helps a little. Once you know which brands the cited pages name, it adds almost nothing.

Mentions on cited pages: 0.59, 93% of batches positive
Links from cited pages: 0.46, 88% of batches positive
Referring domains: 0.19, 66% of batches positive
Organic traffic (US): 0.18, 65% of batches positive
Domain authority: 0.16, 66% of batches positive
Backlinks: 0.16, 66% of batches positive
Organic keywords (US): 0.15, 62% of batches positive

Median rank correlation, inside each batch, between how often the engine names a brand and each measure of that brand, with the share of batches where it is positive. Brands with a known site and an authority figure, 207 to 358 batches per measure.

High authority, no cited page names it: 12.1%, 954 cases
Low authority, no cited page names it: 11.4%, 1,380 cases
Low authority, 2 or more cited pages name it: 48.7%, 95 cases

Share of answers that name the brand. Authority thirds over the 580 brand sites (DataForSEO rank, cut at 228 and 343).

Branded web mentions: 0.664
Branded anchor text: 0.527
Branded search volume: 0.392
Domain Rating: 0.326
Referring domains: 0.295
Branded traffic: 0.274
Backlinks: 0.218

Correlation (Spearman) across 75,000 brands with how often each appears in Google’s AI Overviews, the link measures in gray: Ahrefs, May 26, 2025. Brands with a Domain Rating above 40 only; web mentions include linked ones; bigger brands lead on every measure at once.

One market tracked daily since May

−0.01

correlation between a competitor’s domain authority and how often AI names it, the 54 competitors with authority data, measured only where their own site was cited

The same market, no link

67.4%

of 47,633 namings of a competitor carry no link to its site

Google AI Mode

0.12

authority’s weight beside mentions, the most of any engine; Gemini 0.08, ChatGPT and Claude indistinguishable from zero

n = 580 brands, 5,332 cases, 358 batchesdomain authority: DataForSEO rank, 0 to 1,000, 509 of 537 domains fetched Oct 21, 2026brand sites mostly small and mid-sized: median rank 281, 10 of 537 at 600 or moretraffic and keywords: US estimates, while 36 of 58 reports are UKexplained variance: of the brands’ ranks inside each batch

580 of the 584 brand websites in the test have a domain authority figure, DataForSEO’s link-based rank (the same idea as Ahrefs’ Domain Rating, on a 0 to 1,000 scale). Inside one batch, higher-authority brands are named a little more often: a median correlation of 0.16, positive in 236 batches and negative in 119, with referring domains, backlinks and search traffic at the same level. On the same brands, mentions on the cited pages score 0.59 and predict naming better in 234 batches of 282, and they beat each of the other authority measures by more than three batches to one. Higher-authority brands are somewhat more often named on the cited pages, and most of what authority predicts is shared with mentions: its weight falls from 0.14 alone to 0.05 beside mentions and links, and dropping it from the model loses 0.26 points of explained variance against 10.7 for mentions. A high-authority brand no cited page names is named in 12.1% of answers, about as often as a low-authority brand no page names (11.4%); a low-authority brand that two or more cited pages name, in 48.7%. Where a batch held both, the low-authority brand the pages name was named more often 25 times and the high-authority brand they do not name 8 (5 ties). Authority helps a little, and most of it travels with being written about.

06 / where links still count

Links and rankings track which pages an engine cites. A citation and a mention are different outcomes.

Google says its AI features are rooted in its core search ranking, and in Conductor’s count of 168 million AI Overview citations a page in position 1 is cited 24.9% of the time and one in position 10 10.2%. SE Ranking found referring domains the strongest predictor of which sites ChatGPT cites. In sourcesignal’s data, an engine citing the brand’s own site is the strongest single sign that it names the brand (a median correlation of 0.74), partly because an answer about a brand cites the brand; links keep a weight of their own beside mentions (0.14), and Google AI Mode is the engine where authority holds on most. But the brands AI names most and the sites it cites most overlap less than half the time. Links track the footnote; mentions track the name.

Position 1: 24.9%
Position 2: 21.1%
Position 3: 17.9%
Position 10: 10.2%

How often a page is cited in the AI Overview, by its organic position for the query: Conductor, 168 million AI Overview citations, US, Jun 15 to Aug 15, 2026.

ChatGPT citations

8.4

average for sites with 350,000 or more referring domains, against 1.6 to 1.8 for sites with up to 2,500SE Ranking, Nov 24, 2025

Named and cited

Under half

of the 100 brands AI names most are among the 100 domains it cites most, every month, US and UKSemrush, Oct 13, 2026

Conductor, SE Ranking and Semrush: vendor studies, linked in the MethodConductor and SE Ranking measure citations, not namesown site and links’ weight: sourcesignal’s free-report data

07 / predictable where it counts

Which page an engine cites first is a lottery. Which brands it names tracks the pages.

What moves
Answers led by the most common first link: 46%, median of 386 batches
Next day’s answer with the same first citation: 30.7%, 21,398 pairs, tracked daily
Cited pages cited again 90 or more days later: 16.0%, of 4,517, tracked daily
What holds
Cited answers that include the batch’s top page: 90%, median of 386 batches
Batches where naming rises with mentions: 93.3%, 307 of 329

What moves: the first link inside one batch (free report) and, in the questions tracked daily since May, the first citation from one day’s answer to the next and a page’s return 90 or more days after it first appears. What holds: the batch’s top page, and the batches where naming rises with mentions (a positive rank correlation).

n = 386 batches with 10 or more cited answers; 21,398 pairs of answers tracked dailyfirst link: the first source in the order the engine returned themtracked daily: one answer per question, engine and day

Ask the same question again within a minute and the first link changes: the most common first link leads a median 46% of answers, and 7 different pages take first place in the median batch. In the questions tracked daily since May, the first citation repeats in 30.7% of consecutive answers, and a page cited for a question on an engine is cited for it again 90 or more days later in 16.0% of cases. The pages behind the answer hold steadier than the first link (the top page is in a median 90% of cited answers), and who gets named tracks them: naming rises with mentions in 93.3% of batches, from 8.5% of answers with no naming page to 74.6% at seven or more. First place is a lottery; being named is not.

08 / memory is mentions too

Engines also name brands no cited page mentions. Research finds a model’s recall tracks how often a name was written.

Not every naming comes from a page the engine read. 30.4% of namings happen with no readable cited page naming the brand, an upper bound because some pages could not be read. In the 73 batches where we read pages carrying 80% or more of the citations it is 15.3%, close to the 16% of entities in AI rankings that a University of Toronto team found in no retrieved snippet. Research on how models learn finds recall rises with the number of training documents about an entity (Kandpal et al.) and fails on the least popular ones: about 19% accuracy for one model on the 4,000 least popular questions of one test (Mallen et al.). What a model remembers is built from mentions too.

Our free-report answers
All batches, an upper bound: 30.4%
Batches where we read pages carrying 80% or more of the citations: 15.3%, 73 batches
University of Toronto, Jan 2026
Entities in AI rankings found in no retrieved snippet: 16%, 4 engines

Ours: the share of namings where no cited page we could read names the brand, an upper bound because some pages could not be read. University of Toronto, preprint: the share of entities in generated rankings that appear in no snippet the engine retrieved.

Free ChatGPT searched

14.4%

of 66.6 million logged-out responses, Aug to Sep 2026; the rest came from the model alone, and paid and reasoning models search moreAhrefs, Oct 12, 2026

Six current models

0.63

correlation between a brand’s search interest and how prominently they recommend it, across 82 brands; ad spend and Wikipedia add nothingMalthouse et al., preprint, Sep 14, 2026

n = 449 batches; 73 where the read pages carry 80% or more of the citationsunread pages and skipped Google listings hide some mentionsthe research on memory is not about brands

What it means

Be named on the pages engines read for your buyers’ questions.

Get named where the engine reads: A brand seven or more cited pages name is named in 74.6% of answers; a brand no cited page names, in 8.5%. The pages worth being on are the ones the engines already cite for your buyers’ questions.
Count mentions, linked or not: Two in three mentions on cited pages carry no link, and a link with no mention turned up in 0.7% of cases. Track where you are named, not only where you are linked.
Authority alone did no better: A high-authority brand no cited page names is named in 12.1% of answers, about the same as a low-authority one, and beside mentions authority keeps a weight of 0.05 against 0.44.
The pool, not one page: In the client projects tracked daily, answers citing one third-party page that names the client were not measurably more likely to name it than answers to the same question citing none, on a small sample mostly from two projects (a risk ratio of 1.40, interval 0.79 to 2.48), while across 11 projects the share of cited pages naming the client tracks how often it is named at 0.88. Many pages naming you go with being named; one page alone showed no measurable lift.
Links still go with being cited: Organic position still tracks which pages Google’s AI Overviews cite, and referring domains track which sites ChatGPT cites. A citation puts your page in the answer; your name in it goes with mentions.

sourcesignal asks your buyers’ questions on AI engines, lists every page the answers cite, finds the people behind each page and drafts the pitch to be named there, sent from your own mailbox once you approve it.

Method

Where the answers come from, how they were counted and what they cannot show

Two sets of answers: Every number from sourcesignal’s own data comes from the free report unless it says tracked daily: 58 reports for 49 businesses (36 of the reports for the UK), Aug 7 to Oct 20, 2026. A business enters buying questions and each is asked of up to five engines again and again within about a minute: 576 batches, 17,262 answers. Claude is the API with web search, not the Claude app, and ran on 22 of the 58 reports. Tracked daily means the answers sourcesignal collects for its customers’ projects through Ahrefs Brand Radar: 24,235 answers in 17 projects, May 27 to Oct 21, 2026, one per question, engine and day. No client, business, question or brand is named on this page.
The page fetch: For each batch, up to 15 pages cited in two or more of its answers: 3,179 distinct pages carrying 80.2% of the citations, fetched once on Oct 21, 2026 with a plain request (a desktop browser’s header, no cookies, no login, no scripts run). 2,398 of them (75.4%) could be read: HTML with 500 or more characters of text, not a bot check or a page that only a script fills. They carry 59.4% of all the citations. The rest were blocked (437), empty without scripts (198), behind a bot check (46) or failed in other ways; Google Maps and travel listings (143 rows) were skipped. A plain request cannot see links a script builds, and 36 video, 32 forum and 21 social pages could not be read, so mentions there are missing. The engines read each page between Aug 7 and Oct 20; the fetch reads it as it stood on Oct 21.
Brands: Each report’s business, the competitors it typed in, and the competitors the engines named in two or more of its answers: 944 names, 881 analyzed after names that are domains, everyday words or generic phrases were left out. A brand is named when its name appears as a whole word, in any case, in the answer or on the page; a page on the brand’s own site never counts as naming it. Websites were known for 584 brands, which are the only ones in the link and authority tests. The competitors the engines named are selected on being named, so the businesses and the competitors they typed in, brands nobody selected, are reported as a check.
Domain authority: DataForSEO domain metrics for 580 of the 584 brand sites (537 of 541 domains: 509 fetched Oct 21, 2026 with sourcesignal’s own fetcher, 28 from earlier fetches of the same kind). The rank runs from 0 to 1,000 and is built on links, the same idea as Ahrefs’ Domain Rating (0 to 100) on another scale; referring domains and backlinks are global, organic traffic and keywords are estimates for the US index. One reading of authority is applied to answers from the window before it. The brand sites are mostly small and mid-sized (median rank 281, between 189 and 379 for the middle half).
Measures: Naming share: answers that name the brand, over the batch’s answers. Mentions: the share of the batch’s citations, on pages we could read, that land on pages naming the brand, or for the curve in 02 the number of those pages. Links: the same for pages linking to the brand’s site or a subdomain. Inside one batch: a rank correlation per batch over its brands, tested by the share of batches where it is positive (an exact sign test). Head to head: per batch, which of two measures follows naming more closely on the same brands. Weights: within each batch every measure becomes a rank and is standardized, then one least-squares fit of naming on all of them, with 95% intervals from 2,000 resamples of whole batches; explained variance is of those ranks.
What this shows: Association, not causation. The cited pages are the pages the engine retrieved, as close to a mechanism as these counts get, but retrieval and naming can share a cause: a page about a brand gets cited when an answer discusses that brand. Authority may help a brand get written about, or the reverse; this design cannot tell. The businesses chose to run a free report, most of them in the UK. The one-page effect in What it means was not detected, which is not the same as zero.
Other studies: Ahrefs, Google says links matter less, Jan 30, 2025
Ahrefs, What is a good Domain Rating, Jun 2, 2026
SISTRIX, the US v. Google testimony on ranking (anchors, body, clicks), May 23, 2025
Search Engine Journal, Google’s comments on links and its spam-policy wording, Apr 22, 2024
Google, How Search works: ranking results, read Oct 21, 2026
Ahrefs, AI Overview brand visibility factors, May 26, 2025
Ahrefs, Brand visibility in ChatGPT, AI Mode and AI Overviews, Dec 12, 2025
Google Search Central, Optimizing your website for generative AI features, updated Jul 10, 2026
Conductor, AI Overviews vs organic rankings, updated Sep 25, 2026
SE Ranking, ChatGPT citation factors, Nov 24, 2025
Semrush, AI Visibility Index, H2 2026, Oct 13, 2026
University of Toronto, preprint, arXiv 2601.16858, Jan 23, 2026
Ahrefs, Query fan-out in ChatGPT, Oct 12, 2026
Kandpal et al., ICML 2023, Large language models struggle to learn long-tail knowledge, 2023
Mallen et al., ACL 2023, When not to trust language models, 2023
Malthouse et al., preprint, Evaluating brand retrieval and ranking in LLM recommendations, Sep 14, 2026

Measured October 21, 2026 from the production database, read only, with the page fetch of the same day and domain metrics fetched that day for 509 of 537 domains. Figures describe Aug 7 to Oct 20, 2026 (free report) and May 27 to Oct 21, 2026 (tracked daily); both keep collecting.

Your questions, the same counts

See who gets named for your buyers’ questions.

Most results above start from the pages the engines cite: a brand seven or more of them name was named in three answers in four, and two in three mentions on those pages carried no link. Ask three of your own questions and the free report runs each one again and again on five engines: how often you are named, which competitors are named instead, and the pages each engine cites for the answer.

How the numbers are made